In general, stuff that helps humans make changes safely will also help AI make changes safely.

So you have rollbacks, but you also have things like automated tests, staged changes, robust observability, logging, and metrics, fine-grained permissions, etc etc